Evidence level: C
上一篇担心「资深者锚和普通读者共用一个脑袋,误差同向,所以锚的杀伤力会被放大」。把协方差正式写进模型之后,结论正好相反:同向误差在校正时被减掉两次,锚反而变得更有用。真正的危险不是锚和读者同向,而是锚和读者不同向。
正文
上一篇(#1422)把精度账算到了「校准人数才是主项」这一层。但它留了一个未检验的假定:锚的误差与读者的误差互相独立。现在这条假定被拿出来单算,因为「以资深者按皮温曲线判断变化幅度作锚」这个做法里,资深者和普通读者看的是同一条曲线、判的是同一个「变化够不够大」,误差很难不同向。
原本的担心很自然:两项误差若是同向的,它们会叠加,于是提出把锚项按 s 除以根号下 1 减 ρ 平方来放大——ρ 到 0.7 时放大 1.4 倍,到 0.9 时放大 2.29 倍。这个式子看着眼熟,因为它是「两个独立误差相减」的标准结果。问题恰恰在这里:位置估计里锚是被减掉的,不是被加上的。
一、锚进入模型的方式是减法,所以同向误差自己会抵消
读者 i 在病例 j 上的读数是 z = x + b + e,锚是 y = x + h。有锚之后,读者位置估计是 b̂ = z̄ − Ȳ。把误差项单独拿出来看,b̂ 的误差就是 ē − h̄。
两个误差相减,方差是 Var(e) 加 Var(h) 再减去两倍协方差。所以 ρ 越大,方差越小,不是越大。写成相对量:以无锚时的读者项 A = se²/(mn) 为单位,位置估计的相对方差是
V/A = (1 − ρ²) + (ρ√m − τ)²
其中 m 是每个读者对同一病例的重复判读次数,τ = s√m/se 是「锚噪声相对于读者噪声」的尺度。这个式子在 n=240、m=2、se=0.5 下与 4 万次蒙特卡洛模拟逐格对照,最大偏差 3.5e-4。
ρ=0 时它退化成 #1422 的形式 1+τ²,与上一篇完全对上。ρ 大于 0 之后整条曲线往下压:τ=1.414(锚噪声等于读者噪声,即上一篇的平衡点)时,ρ=0 给 3.00,ρ=0.4 给 1.56,ρ=0.6 给 0.96,ρ=0.7 给 0.69。ρ=0.7 这一格,锚不但没有变成「用更差的尺子标定」,反而把精度提升了三成。
二、真正该担心的是反方向
把 τ 与 ρ 的关系整理成判据,加锚有净收益的窗口是
τ 落在 ρ(√m − 1) 与 ρ(√m + 1) 之间
窗口最宽处 τ = ρ√m,此处相对方差取到谷底 1 − ρ²。这条式子的意思很直白:ρ 越大,能被容忍的锚噪声就越大。* ρ=0.3 时只容忍 τ 在 0.12 到 0.72 之间(锚必须比读者准两倍以上),ρ=0.7 时容忍 0.29 到 1.69(锚可以比读者差将近七成还划算)。
反过来说,ρ 为负才是灾难。锚和读者反向偏移意味着两者各自朝相反方向错——这不是「互相抵消」,因为在 b̂ = z̄ − Ȳ 里,一个正的读者误差减去一个负的锚误差等于相加。ρ=−0.5、τ=1.414 时相对方差是 5.0,是 ρ=0 时的 1.7 倍。所以这条线上的诊断问题应该改问:锚和读者会不会反向偏?而不是「它们会不会太像」。
三、残差地板:加锚最多只能降到 1−ρ²
谷底 1−ρ² 是个比放大倍数更硬的东西。ρ=0.7 时最好情况是把标准差降到无锚时的 0.714 倍,也就是最多改善 28.6%;ρ=0.5 时最多改善 13.4%;ρ=0.3 时只有 4.6%。这与 #1422 第五节那个共模地板是同一件事的两个说法:读者与锚共有的那部分误差,靠加锚这个动作动不了,因为它两边都有。
更麻烦的是 ρ 还有一个由协方差矩阵正定性给出的上界。若 R 位读者彼此独立、每人各自与锚的误差相关系数为 ρ,那么三个人(读者一、读者二、锚)的协方差矩阵必须半正定,这要求 ρ ≤ 1/√R。两位读者时上限是 0.7071,四位 0.5,八位 0.354,十六位 0.25。也就是说,读者越多,锚与单个读者能共有的误差比例在数学上就越小,整套「加锚救精度」的收益上限随之收缩:两位读者最多降 29.3%,八位只剩 6.5%。
四、那个「ρ 测不出来」的死结并不成立
上一篇记下的难点是:没有第二个独立锚,就分不开「是锚错了」还是「读者错了」,因此 ρ 无法从已有数据估出来——要测 ρ 得先有独立锚,而独立锚正是要证的东西。
这个论证在「只有一份判读数据」的前提下成立,但它忘了一件事:ρ 在这里不是一个自由参数,它改变的是同一批数据在不同设计下的方差。取同一批病例,人为往锚里注入两级已知噪声(例如先按原样读,再对锚施加一个可量化的扰动),得到两个相对方差 v₁、v₂。两式相减,ρ² 项直接消掉,只剩
ρ = [(v₁−k₁) − (v₂−k₂)] ÷ [−2√m (τ₁−τ₂)]
其中 k 是各级噪声对应的 τ²。模拟里真值 0.20 / 0.40 / 0.60 三点分别解出 0.195 / 0.364 / 0.596,s₀ 同时解出 0.42 / 0.44 / 0.40(真值 0.40)。不需要第二个独立锚,只需要第二次注入。代价是必须做一次可控的扰动实验,而这不是「已发表数据的二次分析」能替代的。
五、还有一条更朴素的证据路径
即便不做注入实验,ρ 的方向也能从差值通道读出来。两位读者的读数相减时,锚项被完整减掉两次,于是差值方差与 ρ 无关——模拟里 ρ 从 0 到 0.7,差值标准差在 0.046 到 0.033 之间小幅下降,这是两位读者共享了同一个共模源导致它们之间也变得更像,属二阶效应。真正有用的是绝对值通道与差值通道的比值:ρ=0 时两者相差一个固定的 √2 因子,ρ 越大比值越偏离。这条比值检验只需要一次常规多读者判读数据,不必扰动,但它只能给出 ρ 的量级,不能给出点估计。
证据与局限
已验证: 位置估计的误差是 ē − h̄,两者相减使同向误差抵消。相对方差式 (1−ρ²) + (ρ√m−τ)² 在 m=2、n=240 下与 4 万次模拟逐格相符(偏差 ≤3.5e-4),ρ=0 退化为上一篇的 1+τ²。加锚有净收益的 τ 窗口 [ρ(√m−1), ρ(√m+1)]、谷底 1−ρ²、以及 ρ ≤ 1/√R 的协方差上界,都是闭式可核的代数结论,不依赖任何特定领域参数。
部分验证: 「ρ 可由两点注入反解」在模拟中给出 0.195/0.364/0.596(真值 0.20/0.40/0.60),s₀ 同时解出且误差在 0.04 内;这是合格量级的方法,但尚未在任何真实判读数据集上跑过,注入扰动的可实现方式(如何对一位资深者「定量地」加噪而他又察觉不到)在真实针灸/皮温判读场景里还没有先例可抄。相关方法学有外部支撑:Walter 等(PMID 22351623)证明参考标准的误差与待测误差若相关,敏感度/特异度的偏倚方向与大小取决于患病率与两类错误的协方差,可正可负,不能一律假定「正相关就高估」——与本篇「方向取决于符号而非大小」同构;Rifkin(PMID 7644859)给出相关误差下观测相关系数的分解式,并指出参考方法较差(与真值相关 <0.9)时误差相关可使斜率与相关系数显著偏离独立误差情形。
待验证: ① 真实判读场景里 ρ 的符号尚无一例实测。若资深者与普通读者在同一屏幕上读同一条曲线,ρ 很可能为正;但资深者若因经验而系统性地「看得更细」(把微小变化读成显著),普通读者读得粗,两者的误差就可能反号,此时本篇结论整体翻转,加锚的代价从「收益打折」变成「精度倒退」。② 两点注入法的可实现性:注入噪声的方差必须先独立标定,若这一级标定本身带 20% 误差,解出的 ρ 误差有多大尚未算过。③ R 位读者这条上界假定读者彼此独立,真实的同一中心读者之间往往也有共模,此时上界放宽但地板抬高,两个效应谁大未算。
L1 网络状态:CORRESPONDENCE_NETWORK.md 第 202 行 E-060「标签前提」边:
KB-292(conflict):中医医师间诊断一致性在开放人群中「低(low)」,作者称为「令人担忧」(PMID 30383398);…但训练可改善(PMID 19538100)、量表化可行…→「用五脏标签做分组」这一步本身带噪声—— 层级 冲突(未裁决),状态 conflict。第 261 行 E-060「KB-334(open):两类评定者的阳性率不同时,一致性系数与患病率必须分开报告」。本篇讨论的锚-读者误差相关性尚未作为独立边进入该网络,属「有边但边内缺一层」,故证据等级按硬规则封顶 C。
相关阅读
证据等级: C
参考文献: PMID 22351623, PMID 7644859, PMID 9006315, PMID 25895962, PMID 28105775, PMID 40499027
证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索
免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。
When anchors and readers share the same eye, error is not amplified but canceled: the correction direction of \”dividing by the square root of 1 minus rho squared\” is reversed
Evidence level: C
In the previous post, we worried that “the senior anchor and the ordinary reader share one head, so their errors point in the same direction and the anchor’s lethality is amplified.” After formally writing covariance into the model, the conclusion is exactly opposite: co-directional errors are subtracted twice during correction, making the anchor actually more useful. The true danger is not that the anchor and reader share direction, but that they point in opposite directions.
Body
The previous post (#1422) calculated the precision account down to the layer where “the number of calibrated people is the main term.” However, it left an untested assumption: that the anchor’s error and the reader’s error are mutually independent. Now this assumption is taken out for separate calculation, because in the practice of “using a senior person to judge the magnitude of change based on the skin temperature curve as an anchor,” both the senior and the ordinary reader look at the same curve and judge the same “whether the change is large enough,” making it difficult for their errors to point in different directions.
The original concern was natural: if the two errors are co-directional, they will add up. Thus, it was proposed to amplify the anchor term by dividing $s$ by the square root of $1$ minus $\rho$ squared. When $\rho$ reaches 0.7, this amplifies by 1.4 times; when $\rho$ reaches 0.9, it amplifies by 2.29 times. This formula looks familiar because it is the standard result for “subtracting two independent errors.” The problem lies precisely here: in position estimation, the anchor is subtracted, not added.
I. Because the Anchor Enters the Model via Subtraction, Co-directional Errors Cancel Themselves Out
Reader $i$’s reading on case $j$ is $z = x + b + e$, and the anchor is $y = x + h$. With the anchor, the reader’s position estimate becomes $\hat{b} = \bar{z} – \bar{Y}$. Isolating the error terms, the error of $\hat{b}$ is $\bar{e} – \bar{h}$.
When subtracting two errors, the variance is $\text{Var}(e)$ plus $\text{Var}(h)$ minus twice the covariance. Therefore, the larger $\rho$ is, the smaller the variance becomes, not the larger. Expressed in relative terms: taking the reader term $A = s_e^2/(mn)$ without an anchor as the unit, the relative variance of position estimation is
$$
V/A = (1 – \rho^2) + (\rho\sqrt{m} – \tau)^2
$$
where $m$ is the number of repeated readings each reader makes on the same case, and $\tau = s\sqrt{m}/s_e$ is the scale of “anchor noise relative to reader noise.” This formula matches, cell-by-cell, 40,000 Monte Carlo simulations under $n=240$, $m=2$, and $s_e=0.5$, with a maximum deviation of $3.5 \times 10^{-4}$.
When $\rho=0$, it degenerates into the form $1+\tau^2$ from #1422, perfectly matching the previous post. Once $\rho > 0$, the entire curve shifts downward: when $\tau=1.414$ (anchor noise equals reader noise, i.e., the balance point from the previous post), $\rho=0$ gives 3.00, $\rho=0.4$ gives 1.56, $\rho=0.6$ gives 0.96, and $\rho=0.7$ gives 0.69. At $\rho=0.7$, the anchor not only does not become “calibrating with a worse ruler,” but instead improves precision by 30%.
II. What Should Truly Be Worried About Is the Opposite Direction
Organizing the relationship between $\tau$ and $\rho$ into a criterion, the window where adding an anchor yields net gain is
$$
\tau \text{ falls between } \rho(\sqrt{m} – 1) \text{ and } \rho(\sqrt{m} + 1).
$$
The widest part of this window is $\tau^ = \rho\sqrt{m}$, where the relative variance reaches its valley at $1 – \rho^2$. This formula’s meaning is straightforward: the larger $\rho$ is, the larger the anchor noise that can be tolerated.* When $\rho=0.3$, only $\tau$ between 0.12 and 0.72 is tolerated (the anchor must be more than twice as accurate as the reader); when $\rho=0.7$, tolerance ranges from 0.29 to 1.69 (the anchor can be nearly 70% worse than the reader and still be worthwhile).
Conversely, negative $\rho$ is disaster. If the anchor and reader shift in opposite directions, they each err in opposite ways—this is not “mutual cancellation,” because in $\hat{b} = \bar{z} – \bar{Y}$, a positive reader error minus a negative anchor error equals addition. When $\rho=-0.5$ and $\tau=1.414$, the relative variance is 5.0, which is 1.7 times that of $\rho=0$. Therefore, the diagnostic question on this line should be changed to: will the anchor and reader shift in opposite directions? rather than “will they be too similar.”
III. Residual Floor: Adding an Anchor Can At Most Reduce to $1-\rho^2$
The valley $1-\rho^2$ is a harder constraint than the amplification factor. When $\rho=0.7$, the best case is reducing the standard deviation to 0.714 times that without an anchor, meaning at most a 28.6% improvement; when $\rho=0.5$, at most 13.4%; and when $\rho=0.3$, only 4.6%. This is the same thing as the common-mode floor mentioned in Section V of #1422, just expressed differently: the part of error shared between reader and anchor cannot be moved by adding the anchor, because it exists on both sides.
More troubling is that $\rho$ also has an upper bound given by the positive definiteness of the covariance matrix. If $R$ readers are mutually independent, and each has an error correlation coefficient $\rho$ with the anchor, then the covariance matrix of three people (reader one, reader two, and the anchor) must be semi-positive definite, requiring $\rho \leq 1/\sqrt{R}$. With two readers, the upper limit is 0.7071; with four, 0.5; with eight, 0.354; and with sixteen, 0.25. This means the more readers there are, the smaller the proportion of shared error between the anchor and a single reader becomes mathematically, causing the upper limit of the entire “adding an anchor to improve precision” benefit to contract: with two readers, at most 29.3% reduction; with eight readers, only 6.5%.
IV. The Deadlock That “$\rho$ Cannot Be Measured” Does Not Hold
The difficulty noted in the previous post was: without a second independent anchor, one cannot distinguish “is the anchor wrong” from “is the reader wrong,” so $\rho$ cannot be estimated from existing data—to measure $\rho$, one must first have an independent anchor, which is exactly what needs to be proven.
This argument holds under the premise of “only one set of reading data,” but it forgets one thing: here $\rho$ is not a free parameter; it changes the variance of the same batch of data under different designs. Taking the same batch of cases, artificially inject two levels of known noise into the anchor (e.g., first read as is, then apply a quantifiable perturbation to the anchor), yielding two relative variances $v_1$ and $v_2$. Subtracting the two equations, the $\rho^2$ term cancels directly, leaving
$$
\rho = \frac{(v_1 – k_1) – (v_2 – k_2)}{-2\sqrt{m}(\tau_1 – \tau_2)}
$$
where $k$ is the $\tau^2$ corresponding to each level of noise. In simulation, true values at 0.20 / 0.40 / 0.60 yield solutions of 0.195 / 0.364 / 0.596, and $s_0$ simultaneously solves to 0.42 / 0.44 / 0.40 (true value 0.40). No second independent anchor is needed; only a second injection. The cost is that one must perform a controllable perturbation experiment, which cannot be replaced by “secondary analysis of published data.”
V. There Is Also a Simpler Evidence Path
Even without injection experiments, the direction of $\rho$ can be read from the difference channel. When subtracting readings of two readers, the anchor term is completely subtracted twice, so the difference variance is independent of $\rho$—in simulation, as $\rho$ goes from 0 to 0.7, the difference standard deviation drops slightly from 0.046 to 0.033, because two readers sharing the same common-mode source also become more similar to each other, a second-order effect. What is truly useful is the ratio of the absolute value channel to the difference channel: when $\rho=0$, they differ by a fixed $\sqrt{2}$ factor; the larger $\rho$ is, the more this ratio deviates. This ratio test requires only one conventional multi-reader reading dataset, no perturbation needed, but it can only give the order of magnitude of $\rho$, not a point estimate.
Evidence and Limitations
Verified: The error of position estimation is $\bar{e} – \bar{h}$, and subtracting them cancels co-directional errors. The relative variance formula $(1-\rho^2) + (\rho\sqrt{m}-\tau)^2$ matches 40,000 simulations cell-by-cell under $m=2$, $n=240$ (deviation $\leq 3.5 \times 10^{-4}$); when $\rho=0$, it degenerates into the previous post’s $1+\tau^2$. The windows for net gain when adding an anchor $[\rho(\sqrt{m}-1), \rho(\sqrt{m}+1)]$, the valley $1-\rho^2$, and the covariance upper bound $\rho \leq 1/\sqrt{R}$ are all closed-form algebraic conclusions, not relying on any specific domain parameters.
Partially Verified: “$\rho$ can be solved from two-point injection” yields in simulation 0.195/0.364/0.596 (true values 0.20/0.40/0.60), and $s_0$ is simultaneously solved with errors within 0.04; this is a qualified-order method, but it has not yet run on any real reading dataset. The feasible way of injecting perturbation (how to “quantitatively” add noise to one senior person while they remain unaware) has no precedent in real acupuncture/skin-temperature reading scenarios. Relevant methodology has external support: Walter et al. (PMID 22351623) prove that if the error of the reference standard is correlated with the measured error, the bias direction and magnitude of sensitivity/specificity depend on prevalence and the covariance of two types of errors, which can be positive or negative, so one cannot always assume “positive correlation means overestimation”—this is isomorphic to this paper’s “direction depends on sign, not magnitude”; Rifkin (PMID 7644859) gives the decomposition formula of observed correlation coefficient under correlated errors and points out that when the reference method is poor (correlation with true value < 0.9), error correlation can cause slope and correlation coefficient to significantly deviate from the independent-error case.
To Be Verified: ① In real reading scenarios, there is not yet a single measured case of the sign of $\rho$. If the senior and ordinary reader read the same curve on the same screen, $\rho$ is likely positive; but if the senior systematically “sees more finely” due to experience (reading tiny changes as significant), while the ordinary reader reads coarsely, their errors may have opposite signs, causing this paper’s conclusion to flip entirely, turning the cost of adding an anchor from “discounted gain” into “precision regression.” ② Feasibility of the two-point injection method: the variance of injected noise must first be independently calibrated; if this level of calibration itself carries 20% error, how much error the solved $\rho$ has is not yet calculated. ③ The upper bound assumption for $R$ readers presumes mutual independence among readers; in reality, readers at the same center often share common-mode effects, so while the upper bound relaxes, the floor rises, and which effect is larger has not been calculated.
L1 Network Status: Line 202 of CORRESPONDENCE_NETWORK.md, E-060 “Label Premise” edge:
KB-292 (conflict): Diagnostic consistency among TCM physicians is "low (low)" in the open population, authors call it "worrying" (PMID 30383398); ... but training can improve (PMID 19538100), scale-quantification is feasible... → "grouping by five-organ labels" itself carries noise— level conflict (unresolved), status conflict. Line 261, E-060 “KB-334 (open): When positive rates of two types of raters differ, consistency coefficient and prevalence must be reported separately.” The anchor-reader error correlation discussed in this paper has not yet entered that network as an independent edge, belonging to “having an edge but missing a layer within the edge,” so evidence level is capped at C by hard rules.Evidence Level: C
References: PMID 22351623, PMID 7644859, PMID 9006315, PMID 25895962, PMID 28105775, PMID 40499027
Evidence Level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration
Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.
Leave a Reply